USC Debuts Professor Labs AI to Turn Course Curricula Into Live Tutors

From Vision to Reality: Live Demo of Professor Labs’ AI-Powered Learning Platform With USC — Ai4

Ai4

Interactive multi-agent AI allows professors to replace easily pirated textbooks with dynamic, un-piratable course subscriptions while democratizing elite one-on-one tutoring for all students.

AI-assisted summary of From Vision to Reality: Live Demo of Professor Labs’ AI-Powered Learning Platform With USC — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published by the Ai4 channel on YouTube on August 29, 2025, this presentation features USC Senior Lecturer Mike Lee and undergraduate research assistant Manuela Berrio showcasing Professor Labs. According to the description, the session demonstrates a working prototype that transforms traditional university lectures into interactive, gamified, and hyper-personalized learning quests powered by course-specific AI agents.

## TL;DR USC educators Mike Lee and Manuela Berrio present Professor Labs, a multi-agent educational architecture featuring Knowledge AI, Student AI, and Tutor AI designed to align with Bloom's Taxonomy. The platform democratizes round-the-clock personalized tutoring and continuous practice assessments while introducing an interactive curriculum licensing model to replace static, pirated textbooks. It supports local, hybrid, and cloud deployment models with purchasing power parity pricing for global accessibility. In the live demonstration, the presenters illustrate automated 15-week syllabus generation, real-time context injection via audience input, automated lecture slide and exam generation, and high-fidelity multilingual translation.

## Atmosphere & style The presentation is energetic, collegiate, and conversational, combining Silicon Valley evangelism with self-deprecating academic humor about tech salaries and early-morning office hours. The dynamic between the veteran industry-executive professor and his sharp undergraduate researcher feels mutually respectful, candid, and pedagogical. Watching the live demo conveys the spontaneous, authentic tension of real-time software trials, enlivened by lighthearted audience banter and passionate advocacy for human-centered experiential education.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=oMgbGd95lSI&t=4s) Despite widespread information access via AI, the primary value of higher education remains direct human-to-human experiential learning and faculty mentorship. - [03:53](https://www.youtube.com/watch?v=oMgbGd95lSI&t=233s) Professor Labs addresses static curricula and uneven student motivation by mapping personalized study materials and dynamic assessments across Bloom's Taxonomy. - [03:53](https://www.youtube.com/watch?v=oMgbGd95lSI&t=233s) In the generative AI landscape, defensible long-term value resides in building proprietary application layers on top of existing frontier foundation models. - [06:28](https://www.youtube.com/watch?v=oMgbGd95lSI&t=388s) The platform's core architecture divides responsibilities across three agents: Knowledge AI (course IP repository), Student AI (digital twin tracking mastery), and Tutor AI (adaptive explainer). - [06:28](https://www.youtube.com/watch?v=oMgbGd95lSI&t=388s) To align with smartphone-first habits, the system structures curriculum into narrative 'Acts' using gamified micro-quests and personalized reminder personas. - [10:12](https://www.youtube.com/watch?v=oMgbGd95lSI&t=612s) The platform democratizes private one-on-one coaching through 24/7 on-demand tutoring and continuous 'Assess Me' exam question generation. - [10:12](https://www.youtube.com/watch?v=oMgbGd95lSI&t=612s) Content knowledge is decoupled from presentation logic, allowing different specialized frontier models to handle tasks like visual rendering or rapid text Q&A without altering course facts. - [13:10](https://www.youtube.com/watch?v=oMgbGd95lSI&t=790s) Interactive course subscriptions ($10–$20) replace easily pirated static textbooks while adjusting globally to local purchasing power parity. - [17:39](https://www.youtube.com/watch?v=oMgbGd95lSI&t=1059s) The open-source, model-agnostic architecture supports Local, Hybrid, and Cloud deployment configurations to balance data privacy, hardware constraints, and uptime. - [19:22](https://www.youtube.com/watch?v=oMgbGd95lSI&t=1162s) Ingesting a single textbook or slide deck automatically parses content into a full 15-week course syllabus with primary and alternative lecture topics. - [21:17](https://www.youtube.com/watch?v=oMgbGd95lSI&t=1277s) Knowledge AI supports real-time context injection, allowing instructors to immediately update the course knowledge base and have Tutor AI answer queries on new material. - [23:38](https://www.youtube.com/watch?v=oMgbGd95lSI&t=1418s) The platform provides exportable slide decks with speaker notes, high-fidelity multilingual translation, and automated generation of 30-to-40-question midterm exams.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=oMgbGd95lSI&t=4s) Introductions and Background

Mike Lee opens the presentation by introducing himself as a Senior Lecturer of Digital Entrepreneurship and Advanced Computing at the University of Southern California (USC), alongside his undergraduate research assistant, Manuela Berrio. Berrio introduces herself as a rising junior at USC pursuing a dual degree in Computer Science and Business Administration with a specialization in AI applications. Lee commends Berrio as one of his top students who will also serve as a Teaching Assistant for his course on building personal AI agents. He emphasizes that USC students are now actively building custom LLM agents so they can graduate as true "AI natives."

Lee shares his extensive industry background spanning over 20 years, including CTO roles at digital media startups and a technology evangelist position at Netscape. He jokes with the audience about how rare it is for current undergraduate students to even recognize the name Netscape: "Thank God... it's been like five years since a single student knew what that was. So I really appreciate all of you." After spending a decade as a studio Vice President of Technology and building venture technology projects, he chose to take an 80% pay cut to teach at USC. Addressing the educators in the room, he quips that while teachers could easily make five times more in the private sector, they remain in academia because they genuinely care about students. Lee also highlights receiving the student-voted Golden Apple teaching award and serving as the faculty advisor for several innovation initiatives.

Beyond classroom teaching, Lee organizes an annual, fully funded global student startup competition that brings 50 international students from institutions like USC, UC Berkeley, Harvard, Cornell, Oxford, and Aalto University to South Korea to collaborate and compete with 50 local Korean students alongside 20 flown-in faculty mentors. Having just returned from serving as Head of Mentors at the European Innovation Academy with 300 international students, Lee underscores that despite AI making information widely accessible on platforms like YouTube, the irreplaceable value of a university education lies in direct human-to-human experiential learning.

- **Academic Roles**: Senior Lecturer at USC Marshall/Viterbi; founder of global student startup exchange programs. - **Participating Universities**: USC, UC Berkeley (Cal), Harvard, Cornell, Oxford, and Aalto University (Finland). - **Core Philosophy**: Physical universities must pivot toward high-impact experiential learning and mentorship that cannot be replicated by pre-recorded videos or standalone AI tools.

---

[03:53](https://www.youtube.com/watch?v=oMgbGd95lSI&t=233s) The Vision for Professor Labs and Hyper-Personalization

Reflecting on the initial vision introduced at AI4 the previous year, Lee outlines the core structural challenges facing higher education: static, one-size-fits-all curricula and highly uneven student motivation. He candidly describes the typical classroom distribution: one-third of students are self-driven high performers who excel with minimal instruction; one-third leave faculty wondering "how the hell did they get into our university"; and the middle third requires dynamic, tailored motivation to stay engaged. Professor Labs was created to bridge this divide through hyper-personalized learning and gamification.

The platform grounds its progression in Bloom's Taxonomy, mapping foundational understanding (rote memorization and core definitions) to the first half of a semester, followed by application, analysis, and active classroom engagement in the second half. Lee notes that as underlying Large Language Models improve, educational software can infinitely generate tailored study guides, exercises, and assessments at every tier of Bloom's hierarchy. Furthermore, Lee points out that unless an organization is spending tens of millions of dollars a month training frontier models like OpenAI or Google, the defensible value lies entirely in the application layer built on top of foundational models.

```mermaid flowchart TD accTitle: Bloom's Taxonomy Integration in Professor Labs subgraph Progression ["Bloom's Taxonomy Application"] A["Higher-Order: Analyze, Evaluate & Create"] --> B["Intermediate: Apply & Comprehend"] B --> C["Foundations: Remember & Define"] end subgraph Execution ["Platform Capability"] D["Infinite Adaptive Case Studies & Syntheses"] E["Dynamic Interactive Problem Sets"] F["Automated Rote Drills & Flashcards"] end C --> F B --> E A --> D ```

---

[06:28](https://www.youtube.com/watch?v=oMgbGd95lSI&t=388s) The Professor Labs AI Architecture: Knowledge, Student, and Tutor Agents

Lee breaks down the proprietary multi-agent framework powering Professor Labs, which centers on three interconnected AI engines:

```mermaid flowchart TD accTitle: Professor Labs Multi-Agent Core Framework subgraph Inputs ["Instructor Content"] KB["Textbooks, Slide Decks, Papers, Syllabus"] end subgraph CoreEngine ["Professor Labs Engine"] KAI["Knowledge AI\n(Course Brain / Professor's IP)"] SAI["Student AI\n(Digital Twin / Competency Tracker)"] TAI["Tutor AI\n(Adaptive Explainer / Socratic Guide)"] end subgraph Experience ["Student Experience"] GAME["Gamified Mobile Micro-Quests"] VOICE["Adaptive Persona / Tone Engine"] end KB --> KAI KAI <--> SAI KAI --> TAI SAI --> TAI TAI --> GAME TAI --> VOICE ```

The **Knowledge AI** serves as the repository for the instructor's intellectual property. A professor can upload full textbooks, recent research papers, updated slide decks, or lecture transcripts directly into the Knowledge AI, establishing an accurate, course-specific "brain." The **Student AI** operates as a digital twin tracking individual student mastery across every topic (e.g., scoring 20% in module four but 90% in module five), allowing the system to identify knowledge gaps relative to target learning outcomes.

The **Tutor AI** dynamically formats explanations to fit individual learning preferences—whether a student needs an intuitive visual breakdown, an audio summary, or a simplified explanation "as if they're a fifth grader." Lee compares traditional teaching to parenting: students often tune out their own professors while praising an outside guest lecturer who delivers the exact same message. Tutor AI circumvents this dynamic by acting as an objective, customizable study partner.

To boost engagement, the platform integrates gamified micro-learning tailored to smartphone-first habits. Lee illustrates the generational shift with an anecdote about the global K-pop group KATSEYE: when asked to act out the word "phone," older generations mimic holding a telephone handset to their ear, whereas younger members pantomime looking down at a flat rectangular screen. Professor Labs leverages these mobile micro-habits for rote drills and uses adaptive notification personalities reminiscent of Duolingo's persistent, meme-worthy reminders. Course curricula are structured into narrative "Acts" (aligned with midterms) containing individual lecture modules.

---

[10:12](https://www.youtube.com/watch?v=oMgbGd95lSI&t=612s) Pedagogical Methodology: Continuous Assessment and Multimodal Learning

The primary pedagogical objective of Professor Labs is reinforcing and scaling in-person instruction rather than replacing it. Lee contrasts affluent students who grew up with $100 to $200-per-hour private tutors against students whose busy parents could not provide academic coaching. Professor Labs aims to democratize this elite one-on-one tutoring experience for every enrolled student.

The platform provides 24/7 on-demand support through two dedicated modes: a **Private Tutor** that guides students through difficult concepts according to their learning style, and an **Assess Me** feature that continuously generates infinite variations of practice exam questions. This provides round-the-clock help regardless of erratic student study hours:

> "At any time, they could ask this 24/7. Doesn't matter if your TAs don't respond to discussion boards... because students have wacky schedules—midnight, whichever time, etc. Never at 8:00 AM. That's why I make all my office hours 8:00 AM. You must really want to see me."

Lee also shares candid lessons from experimental features, noting that attempting to prompt frontier models to generate interactive vector visualizations (SVGs) of complex physical systems (such as wind turbines) frequently failed in practice. Consequently, Professor Labs decouples content knowledge from presentation, routing tasks to specialized frontier models (like GPT-5/advanced multimodal models for visual rendering and GPT-4 for rapid text Q&A) while keeping the foundational course knowledge constant.

---

[13:10](https://www.youtube.com/watch?v=oMgbGd95lSI&t=790s) Academic Monetization, Democratization, and Platform Architecture

Lee addresses a systemic flaw in academic publishing: professors currently have little incentive to produce dynamic digital materials because textbooks costing hundreds of dollars are routinely pirated into free PDFs. Professor Labs introduces an interactive curriculum licensing model where professors license their Knowledge AI directly to students as required course materials for $10 to $20 per semester. Because the platform offers interactive tutoring and live dynamic assessments, it cannot be pirated as a static PDF.

```mermaid flowchart LR accTitle: Professor Workflow and Licensing Pipeline A["Upload Course IP\n(Slides, Books, Transcripts)"] --> B["Knowledge AI Generates\nStructured Syllabus & Acts"] B --> C["Instructor Refines\n& Customizes Content"] C --> D["Publish Interactive Course\n($10-$20 Student Subscription)"] ```

Lee also focuses on the democratization of AI for faculty and international students. While top-tier US universities have faculty who are technical super-users, instructors worldwide need turnkey tools that abstract technical complexities. Furthermore, student access pricing must adjust dynamically to local economies via purchasing power parity—scaling down from a $50 baseline in the United States to roughly $1.59 in India based on relative GDP per capita.

---

[17:39](https://www.youtube.com/watch?v=oMgbGd95lSI&t=1059s) Deployment Models and Setting Up the Live Demo

To ensure data sovereignty and flexibility, Professor Labs is designed as an open-source, AI platform-agnostic architecture evaluated across three deployment configurations:

| Deployment Mode | Infrastructure & Processing | Key Advantages | Trade-offs & Limitations | |:--- |:--- |:--- |:--- | | **Local** | Runs entirely on a local PC/Mac with local LLMs | Complete data privacy; operates offline without internet | Requires high-end GPU hardware; smaller models lack frontier reasoning | | **Hybrid** *(Demoed)* | Local data storage paired with cloud LLM APIs (e.g., OpenAI) | Keeps source files local while utilizing advanced cloud reasoning | No 24/7 student access when host machine is offline | | **Cloud** | Fully hosted cloud infrastructure with enterprise zero-retention APIs | High availability; 24/7 student access; zero model training on user data | Ongoing cloud compute and API operating costs |

Lee hands the presentation over to Manuela Berrio to run the live demonstration, which will showcase the two-way interaction between the instructor-facing **Knowledge AI** ingestion pipeline and the student-facing **Student/Tutor AI** interface.

[19:22](https://www.youtube.com/watch?v=oMgbGd95lSI&t=1162s) Course Generation and the Professor Labs Interface

The demonstration transitions into the functional platform, illustrating the dual-sided nature of education where faculty interact with one interface and students interact with another. To set up a full 15-week course syllabus, the instructor begins by designating the midterm week—typically set at Week 7 as the midpoint of the term. The professor then uploads course materials, such as a textbook PDF, presentation slides, or reading packet. In this live demo, Professor Mike Lee’s book is uploaded, prompting the underlying large language model to parse and chunk the material into a structured schedule with class meetings mapped out for Mondays and Wednesdays.

The platform architecture centers around two primary pillars: **Knowledge AI** on the instructor side, which allows professors to inject hyper-personalized, topic-specific notes and supplemental slides beyond the baseline textbook, and **Tutor AI** on the student side, designed to serve as an intelligent, round-the-clock teaching assistant. While the production roadmap moves toward a unified cloud-native environment, the current prototype displays these components side by side. Instructors can view extracted lecture topics for each week alongside AI-suggested alternative topics, enabling manual overrides if an educator finds a different concept more relevant to their curriculum.

```mermaid flowchart TD accTitle: Professor Labs Course Generation Workflow accDescription: Flowchart illustrating how course materials are ingested into Professor Labs to generate syllabi, lecture notes, and student interactions.

subgraph Ingestion["1. Ingestion & Setup"] A["Upload Material<br/>(Book / PDF / Slides)"] --> B["Set Midterm Week<br/>(e.g., Week 7)"] B --> C["LLM Chunking Engine"] end

subgraph CourseStructure["2. Course Structuring"] C --> D["15-Week Syllabus<br/>(Mon / Wed Schedule)"] D --> E["Core & Alternative Topics"] end

subgraph KnowledgeAndTutor["3. Personalization & Delivery"] E --> F["Knowledge AI<br/>(Instructor Custom Content)"] F --> G["Tutor AI<br/>(Personalized Student Tutor)"] F --> H["Topic Notes & Slide Decks"] F --> I["Quizzes & Midterm Generator"] end ```

- **Syllabus Automation**: Automatically structures 15 weeks of coursework from a single uploaded document. - **Flexible Topic Selection**: Displays primary extracted topics alongside alternative recommendations for faculty curation. - **Knowledge AI vs. Tutor AI**: Knowledge AI handles instructor-guided context enrichment, while Tutor AI serves personalized student queries.

---

[21:17](https://www.youtube.com/watch?v=oMgbGd95lSI&t=1277s) Live Audience Test with Knowledge AI and Tutor AI

To prove that the platform dynamically integrates new context in real time rather than relying on pre-trained canned responses, the presenters invite the audience to provide an arbitrary, out-of-domain personal fact that could not be found anywhere in Professor Lee's book. Audience members playfully joke about volunteering sensitive details like "social security numbers" or maiden names, before an attendee named Alex volunteers a personal historical anchor: *"When Mount St. Helens blew in 1980, I was five years old."*

The team types this fact into a document, saves it, and uploads it directly into the Knowledge AI module for a specific topic module. The goal is to show that Tutor AI immediately inherits this newly added context and can answer specific factual questions about Alex’s age and the 1980 eruption. As with live software demonstrations, the presenters navigate file saving, local downloads, and uploading constraints in real time before pivoting to showcase topic note generation.

- **Dynamic Context Injection**: Allows real-time ingestion of supplemental text to augment the primary course model. - **Audience Fact**: Alex was 5 years old when Mount St. Helens erupted in 1980.

---

[23:38](https://www.youtube.com/watch?v=oMgbGd95lSI&t=1418s) Topic Notes, Multilingual Capabilities, and Assessment Generation

Moving into the core instructional tools, the presenters demonstrate the automatic generation of comprehensive topic notes. When an instructor triggers this feature, the system generates a ready-to-use lecture outline structured for direct export into presentation software such as Google Slides or Microsoft PowerPoint. Each generated deck includes slide-by-slide content guidelines, detailed speaker notes for the lecturer, and a structured 20-topic breakdown of the selected subject area.

The platform also integrates multilingual learning features designed to assist non-native English speakers. Drawing from her own background growing up in Colombia until age 17, Manuela Berrio highlights how native-language AI support removes comprehension barriers for international students. The translation engine delivers high-fidelity curriculum translations into languages such as Spanish, Korean, and German without degrading technical accuracy. Additionally, the system features interactive gamification modules, automatically creating mini-game quizzes and building comprehensive 30-to-40-question midterm exams containing both standard multiple-choice and complex application-based problem sets.

- **Lecture Slide Generation**: Produces full slide breakdowns with detailed speaker notes and 20-topic conceptual splits. - **Multilingual Learning**: Leverages advanced translation models across languages including Spanish, Korean, and German to support global and ESL learners. - **Assessment Engine**: Generates gamified mini-quizzes and 30–40 question midterm exams combining multiple-choice and scenario-based application questions.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=oMgbGd95lSI&t=4s) Thank God... it's been like five years since a single student knew what that was. So I really appreciate all of you. — Mike Lee

> [03:53](https://www.youtube.com/watch?v=oMgbGd95lSI&t=233s) how the hell did they get into our university — Mike Lee

> [06:28](https://www.youtube.com/watch?v=oMgbGd95lSI&t=388s) as if they're a fifth grader — Mike Lee

> [10:12](https://www.youtube.com/watch?v=oMgbGd95lSI&t=612s) At any time, they could ask this 24/7. Doesn't matter if your TAs don't respond to discussion boards... because students have wacky schedules—midnight, whichever time, etc. Never at 8:00 AM. That's why I make all my office hours 8:00 AM. You must really want to see me. — Mike Lee

> [21:17](https://www.youtube.com/watch?v=oMgbGd95lSI&t=1277s) When Mount St. Helens blew in 1980, I was five years old. — Alex

## Mentioned Mike Lee, Manuela Berrio, University of Southern California (USC), USC Marshall School of Business, USC Viterbi School of Engineering, Professor Labs, AI4, Netscape, UC Berkeley, Harvard University, Cornell University, University of Oxford, Aalto University, European Innovation Academy, Bloom's Taxonomy, OpenAI, Google, KATSEYE, Duolingo, GPT-4, GPT-5, Google Slides, Microsoft PowerPoint

## Video description Mike Lee, Senior Lecturer of Digital Entrepreneurship & Advanced Computing, University of Southern California Manuela Berrio, Undergraduate Research Assistant, University of Southern California

Abstracts

At AI4 2024, we introduced Professor Labs as an academic project aimed at transforming higher education through generative AI and gamified learning. One year later, we return with a working prototype and a live demonstration.

In this session, Mike Lee (Senior Lecturer of Digital Entrepreneurship & Advanced Computing, University of Southern California) and Manuela Berrio (Computer Science & Business Administration major, University of Southern California) showcase how Professor Labs enables professors to transform traditional lectures into interactive, AI-powered learning quests. Built by and for educators, the platform leverages custom large language models trained on course-specific content to deliver hyper-personalized tutoring, continuous assessment, and student-led gameplay.

This live demo will highlight: - The Professor Use Case: Uploading course content, generating assessments, and guiding pedagogy - The Tutor AI: Real-time, personalized feedback and support tailored to individual students

Join us to see how academic research is translating vision into practice—and shaping the future of teaching and learning.

[Source](https://www.youtube.com/watch?v=oMgbGd95lSI)

中文简报

南加大推出 Professor Labs AI 将课程大纲转化为实时导师 交互式多智能体 AI 助力教授以动态、防盗版的课程订阅替代易遭盗版的教科书，同时让所有学生均能享受精英级的一对一辅导。 南加州大学（USC）教育工作者 Mike Lee 和 Manuela Berrio 展示了 Professor Labs，这是一种包含知识 AI（Knowledge AI）、学生 AI（Student AI）和导师 AI（Tutor AI）的多智能体教育架构，旨在契合布鲁姆认知分类学（Bloom's Taxonomy）。该平台不仅普及了全天候个性化辅导和持续练习评估，还引入了交互式课程授权模式来替代静态且易被盗版的教科书。它支持本地、混合和云端部署模式，并采用购买力平价定价以提升全球可及性。在现场演示中，主讲人展示了自动化生成15周教学大纲、通过听众输入实时注入上下文、自动生成讲义幻灯片与考试，以及高保真多语言翻译等功能。 南加州大学（USC）讲师 Mike Lee 与研究助理 Manuela Berrio 现场展示了 Professor Labs，该平台利用定制大语言模型将传统静态课程转换为游戏化的互动学习任务。 (0:04) 平台采用三层 AI 智能体架构：Knowledge AI 负责存储教师的课程知识产权，Student AI 作为数字孪生追踪学生掌握度，Tutor AI 则根据不同学习偏好提供自适应辅导与全天候答疑。 (6:28) 为契合移动端习惯并降低成本，课程内容被组织为叙事化微任务，并支持按购买力平价设定每学期 10 至 20 美元的互动订阅，替代易被盗版的静态 PDF 教材。 (10:12) 现场演示展示了教师端一键将教材拆解为 15 周教学大纲与幻灯片讲稿，并在实时注入外部新事实后，Tutor AI 可立即吸收上下文进行问答及跨语言评估生成。 (19:22)

Reading summary

Mike Lee and Manuela Berrio of the University of Southern California presented Professor Labs, an open-source, multi-agent AI platform designed to automate course assembly, personalize tutoring, and provide continuous student assessment. Lee argues that as frontier models advance, defensible educational value lies in the application layer, which can democratize elite one-on-one coaching while reserving physical classrooms for human mentorship and experiential learning. Multi-Agent Architecture Grounded in Bloom's Taxonomy (6:28) Professor Labs operates on a proprietary triad of specialized AI engines built atop foundational large language models. The instructor-facing Knowledge AI acts as the central course repository, ingesting textbooks, syllabi, slide decks, and lecture transcripts to establish a course-specific knowledge base. A Student AI functions as a digital twin tracking individual mastery across individual modules, while a Tutor AI dynamically formats explanations to fit individual learning preferences, ranging from visual breakdowns to simplified analogies. The system maps curricular progression directly to Bloom's Taxonomy. Foundational tiers, including rote memorization and core definitions, are drilled during the initial half of a term, while intermediate and higher-order tiers—comprehension, application, and synthesis—are supported through dynamically generated case studies and interactive problem sets. Lee asserts that because training frontier foundation models requires tens of millions of dollars per month, educational innovation and defensibility reside almost entirely within this application layer. Automated Syllabus Generation and Dynamic Ingestion (19:22) During live demonstrations, the presenters illustrated how uploading primary course materials—such as an instructor's textbook PDF—prompts the underlying language model to chunk content into a complete 15-week syllabus structured around a designated midterm week. Instructors can inspect extracted lecture topics alongside AI-recommended alternative topics, maintaining manual oversight to curate or override suggested content to match their syllabus. The framework also supports real-time contextual updates. In an audience test, presenters uploaded an arbitrary external fact—that an attendee was five years old when Mount St. Helens erupted in 1980—into a specific module, enabling the student-facing Tutor AI to answer queries incorporating that newly supplied detail immediately. Additionally, the system generates ready-to-use lecture slides complete with speaker notes, supports native-language translations across languages including Spanish, German, and Korean, and automatically compiles 30-to-40-question midterms combining multiple-choice and applied problem sets. 24/7 Personalized Coaching and Continuous Assessment (10:12) The pedagogical objective of Professor Labs is reinforcing and scaling in-person instruction rather than replacing physical classrooms, which Lee argues remain irreplaceable for collaborative, human-to-human experiential learning. Lee highlights an equity gap between affluent students who can afford $100 to $200 per hour private tutors and students without access to supplemental coaching. The platform attempts to bridge this disparity through round-the-clock "Private Tutor" and "Assess Me" features, accommodating irregular student study hours when human teaching assistants are unavailable. To maintain engagement across varying student motivation levels, the platform incorporates mobile-first micro-learning habits, adaptive notification personalities, and gamified quizzes. To ensure reliable multimodal presentation, the architecture decouples core subject knowledge from presentation formatting. Specialized models are routed dynamically—using tools like GPT-4 for rapid text Q&A and advanced multimodal models for visual generation—after initial attempts to generate interactive vector visualizations directly through prompt engineering proved unreliable. Deployment Trade-offs and Global Economic Access (13:10) Professor Labs is designed as an open-source, AI platform-agnostic system evaluated across three deployment configurations: local, hybrid, and fully cloud-hosted. A local deployment runs offline on personal hardware with local LLMs, preserving strict data privacy at the expense of high GPU hardware demands and reduced model reasoning capability. The hybrid model keeps source documents local while utilizing cloud reasoning APIs, and the cloud configuration provides continuous availability under enterprise zero-retention data agreements at the expense of ongoing API operational costs. To counter textbook piracy and provide faculty with incentives to produce digital materials, Lee proposes an interactive curriculum licensing model where students pay a $10 to $20 semester subscription directly for the interactive Knowledge AI tool. To broaden international access, Lee advocates adjusting student subscription fees through purchasing power parity, scaling down from a $50 United States baseline to roughly $1.59 in India based on relative GDP per capita. While Professor Labs illustrates how multi-agent architectures can automate course delivery, personalize tutoring, and bypass static textbook piracy, practical trade-offs remain—including the GPU demands of purely local execution and technical hurdles in automated visual rendering. The broader educational impact also depends on whether academic institutions and global students adopt interactive per-course software subscriptions at scale.

精炼阅读摘要

南加州大学讲师 Mike Lee 与研究助理 Manuela Berrio 介绍了教育平台 Professor Labs，主张高校应依托多智能体架构实现超个性化学习与自动化课件生成，将大学的核心价值从单向信息传授重新锚定在人际体验式教学上。 教育痛点与布鲁姆层级重构 (3:53) Mike Lee 指出，传统高等教育长期受制于“一刀切”的静态课程体系以及学生学习动机严重分化的困境。他观察到典型的大学课堂通常呈现三分结构：三分之一的学生自主性强且成绩拔尖，三分之一的学生缺乏基础与准备，剩余三分之一则极需动态激励以维持专注。鉴于基础知识在 YouTube 等网络平台上已高度普及，Lee 强调物理大学不可替代的根本价值在于高影响力的人际体验式学习与导师指导，而非单纯灌输理论概念。 为此，Professor Labs 依据布鲁姆认知层级（Bloom's Taxonomy）推进教学安排：学期前半段聚焦于底层的机械记忆与核心概念定义，后半段则过渡到高阶的应用、分析与主动课堂实践。Lee 认为，除非每月耗费数千万美元训练前沿基础模型，否则教育机构的技术防线完全在于顶层应用层，即利用基础模型在布鲁姆层级各阶段无限生成自适应案例研究、互动问题集与复习卡片。 知识学生与导师三智能体系统 (6:28) Professor Labs 的专有架构由三个互相协作的 AI 引擎构成。面向教师的“知识智能体”（Knowledge AI）作为课程知识产权库，可直接摄取整本教材、最新学术论文、幻灯片与讲义文本，建立课程专属的知识中枢。充当数字孪生的“学生智能体”（Student AI）持续追踪个体在各模块的掌握情况（例如记录学生在第四模块得分 20% 但在第五模块达到 90%），借此动态识别其与既定学习目标间的差距。 “导师智能体”（Tutor AI）则负责根据学生的认知习惯调整表达方式，包括提供直观的视觉图解、音频摘要或通俗简明的解释。Lee 认为传统师生互动常伴随类似亲子关系的沟通倦怠，而 Tutor AI 能作为客观、可定制的学伴打破这一阻碍。该平台还顺应智能手机操作习惯融入移动微学习与游戏化机制，按照期中期末节奏将课程编排为叙事化的“幕”（Acts），并借助拟人化通知提醒提升学生学习黏性。 辅导平民化与多模态模型解耦 (10:12) 该平台的核心教学目标在于强化和扩展线下教学，而非将其取代。Lee 对比了高收入家庭可支付每小时 100 至 200 美元的私人辅导、而工薪家庭学生缺乏课外学术指导的现实，提出借助技术实现精英级一对一辅导的平民化。平台提供两类全天候功能：“私人导师”（Private Tutor）依照学生特定习惯拆解难点，“评估我”（Assess Me）则适应学生极不规律的作息（如深夜复习），无限生成练习题变体，打破了助教答疑时限的限制。 在底层架构上，团队吸取了前期试验的教训：直接提示前沿大模型生成复杂物理系统（如风力涡轮机）的交互式矢量图（SVG）往往频繁失败。因此，平台采取了知识内容与呈现形式解耦的设计，针对不同任务分流路由至专业模型，例如调用 GPT-5 等前沿多模态模型负责图像渲染，调用 GPT-4 负责高速文本问答，同时保持底层课程知识中枢一致。 交互许可机制与灵活部署架构 (13:10) Lee 直指传统学术出版模式的系统性弊端：售价数百美元的纸质教材极易被盗版为免费 PDF 传播，导致教授缺乏制作动态数字化教材的动力。Professor Labs 提出了一种交互式课程许可模式，教授将自身的 Knowledge AI 直接授权给学生作为必选教材，每生每学期仅收费 10 至 20 美元。由于系统内置实时动态辅导与互动测试，该形式无法被静态 PDF 替代。此外，平台基于各国人均 GDP 引入购买力平价动态定价，使美国本土基准为 50 美元的课程在印度可降至约 1.59 美元。 为保障院校数据主权与灵活性，平台设计为中立开放架构，并支持三类部署方案：本地模式完全在个人电脑或 Mac 上运行本地开源模型，离线可用且数据隐私绝对封闭，但依赖高端硬件且推理能力受限；混合模式在本地保存源文件的同时调用云端推理 API；云托管模式则依托企业级零数据保留 API 提供 24 小时高可用服务，但伴随持续的计算与 API 调用成本。 自动化课件生成与动态实测 (19:22) 在现场演示中，Manuela Berrio 展示了通过上传单份教材 PDF 自动生成 15 周大纲的流程。教师设定第 7 周为期中后，系统自动将内容切片为匹配周一与周三排期的教学日程，并列出核心主题与供教师手动微调的备选建议。系统还能一键导出包含逐页内容、详细演讲者备注和 20 个概念拆分的讲义大纲，并直接兼容 Google Slides 或 PowerPoint。 为证实平台具备实时动态学习能力而非依赖预设答案，演示现场将听众 Alex 提出的冷门个人事实（“1980年圣海伦斯火山爆发时我五岁”）录入并上传至特定知识模块，Tutor AI 随即能基于该事实精准回答相关提问。此外，平台展示了针对西班牙语、韩语和德语的高保真专业翻译能力，并演示了可生成 30 至 40 道包含基础选择题与场景应用题的期中考试构建工具。 Professor Labs 展示了多智能体协作重塑大学课程开发、个性化答疑与教材发行模式的潜力。然而，现场演示亦反映出本地文件处理流程较繁琐、复杂矢量图渲染依赖模型分流等工程限制；其倡导的低价交互订阅制与购买力平价模型，在传统高校采买体系与跨国教育市场中的商业可行性仍待进一步检验。

Detailed notes — zh

## Background 该视频由 Ai4 于 2025 年 8 月 29 日在 YouTube 平台发布。视频记录了南加州大学（USC）高级讲师 Mike Lee 与本科生研究助理 Manuela Berrio 的专题演讲与现场实机演示，展示了他们针对高等教育痛点打造的 AI 驱动游戏化学习平台 Professor Labs，旨在将传统单向讲座转化为具备超个性化辅导与自适应评估的学习体验。 ## TL;DR 演讲系统介绍了专为大学师生研发的 Professor Labs 平台，致力于借助生成式 AI 解决千篇一律的课程设置与学生动力失衡等顽疾。平台基于布鲁姆教育目标分类法，构建了涵盖 Knowledge AI、Student AI、Tutor AI 及游戏化机制的多智能体解耦架构，支持教授上传课件一键生成 15 周教学大纲、逐页演讲备注与期中考卷。此外，团队不仅提出了重构教材知识产权的低成本课程订阅与购买力平价定价模型，还现场实机演示了课程知识挂载、即时提取与跨语言教学支持的落地能力。 ## Atmosphere & style 现场氛围既展现了学术研讨的深度，又洋溢着科技初创团队路演般的极客活力。讲者 Mike Lee 凭借资深业界背景幽默风趣地调侃学生作息与传统大学教学现状，与沉稳专注、代表国际学生视角的 Manuela Berrio 形成了鲜明互补的双人互动节奏。演示环节虽偶遇现场技术小插曲，但讲者以真实坦率的台风化解，整体传递出一种致力于打破高额私教壁垒、推动教育平权的真诚与热忱。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=oMgbGd95lSI&t=4s) 在 AI 和网络资源普及使单向知识传授贬值的时代，大学不可替代的真正护城河在于面对面的体验式连接与实际项目实践。 - [03:53](https://www.youtube.com/watch?v=oMgbGd95lSI&t=233s) Professor Labs 致力于通过生成式 AI 与游戏化机制解决高等教育中静态课程、动力不均和缺乏超个性化辅导的三大痛点。 - [03:53](https://www.youtube.com/watch?v=oMgbGd95lSI&t=233s) 系统基于布鲁姆教育目标分类法构建底层教学逻辑，利用不断进化的 LLM 实现从基础记忆到高阶应用与分析的自适应教学。 - [03:53](https://www.youtube.com/watch?v=oMgbGd95lSI&t=233s) 初创企业不应盲目在底层大语言模型上与巨头竞争，核心竞争力应深耕于基础模型之上的垂直应用与教学框架层。 - [06:28](https://www.youtube.com/watch?v=oMgbGd95lSI&t=388s) Professor Labs 采用 Knowledge AI、Student AI 与 Tutor AI 四位一体的架构，分别承载教授知识库、学生掌握度画像与自适应教学。 - [09:56](https://www.youtube.com/watch?v=oMgbGd95lSI&t=596s) 平台通过 24/7 待命的 AI 导师打破高昂私人家教的资源壁垒，实现教育平权并支持基于课程的无限自适应评估。 - [11:25](https://www.youtube.com/watch?v=oMgbGd95lSI&t=685s) 系统采取应用层与底层模型解耦的架构，保持核心知识库固定，根据成本和能力动态调用不同模型来呈现多模态内容。 - [13:08](https://www.youtube.com/watch?v=oMgbGd95lSI&t=788s) 平台正在将纯粹的备课工具升级为互动课程订阅平台，使教授能够将专有知识资产直接转化为可持续的许可收益。 - [13:08](https://www.youtube.com/watch?v=oMgbGd95lSI&t=788s) 为推动全球教育普惠，平台提出了基于购买力平价的动态定价机制，显著降低发展中国家学生的使用门槛。 - [16:18](https://www.youtube.com/watch?v=oMgbGd95lSI&t=978s) 系统评估了纯本地、混合以及云端三种部署模式，其中云端部署被视为支持全天候访问与数据安全承诺的最优形态。 - [19:12](https://www.youtube.com/watch?v=oMgbGd95lSI&t=1152s) 现场实机演示了上传教材 PDF 即可一键自动生成 15 周结构化教学大纲，并支持动态外挂补充材料与实时问答检索。 - [23:38](https://www.youtube.com/watch?v=oMgbGd95lSI&t=1418s) 系统可自动生成包含逐页规划、讲者备注及深度知识点拆解的主题讲义大纲，大幅减轻教师备课负担。 - [24:11](https://www.youtube.com/watch?v=oMgbGd95lSI&t=1451s) 平台具备多语言学术翻译、轻量互动小测验游戏化模块，以及一键生成包含多题型的完整期中试卷等扩展功能。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=oMgbGd95lSI&t=4s) 讲者背景与大学教育的独特价值 南加州大学（USC）数字创业与应用计算高级讲师 Mike Lee 与其本科生研究助理 Manuela Berrio（主修计算机科学与工商管理，专注于 AI 应用）共同登台。Mike 介绍道，Manuela 是他的优秀学生之一，不仅参与该领域的研究，还将担任助教，指导学生构建个性化智能体（Personal AI Agents）。如今他的课程要求所有学生都亲自动手开发个人 AI 智能体与大语言模型应用，让他们在毕业时真正成为“AI 原住民”（AI Natives）。 Mike 分享了自己跨界学术圈的经历：他曾拥有超过 20 年的从业经验，在多家媒体公司与初创企业担任首席技术官（CTO）；早期还在 Netscape 担任技术布道师（Technology Evangelist），并在好莱坞影视制片厂担任了 10 年的技术副总裁，主持过大量前沿科技的创投孵化。后来他放弃了业界数倍的高薪全职投身 USC 执教，并凭借极佳的教学口碑赢得了由学生投票评选出的“金苹果奖”（Golden Apple Award）。 在探讨大学教育的未来时，Mike 强调，在 AI 和 YouTube 泛滥的时代，传统知识传授已被极大地去中心化，大学不可替代的真正价值在于“面对面的体验式连接”（Experiential Learning）。为此，他创办了全球学生创业大赛（Global Student Startup Competition），全额资助来自 USC、UC Berkeley（Cal）、哈佛大学、康奈尔大学、牛津大学以及芬兰阿尔托大学（Aalto University）等名校的 50 余名国际学生飞赴韩国，与当地 50 名学生同台竞技，并配备 20 名业界导师与教授进行深度辅导。同时，作为欧洲创新学院（European Innovation Academy）的导师主管，他观察到全球各地的学生如今都能极快地利用 AI 智能体完成产品构建与敏捷迭代。 ```mermaid flowchart TD accTitle: The Experiential Value of Higher Education in the AI Era subgraph Challenge["知识普及与技术冲击"] A["LLMs / 生成式 AI"] B["YouTube / 在线开源课程"] end subgraph Solution["大学的核心护城河：无法复制的体验"] C["面对面深度研讨与社群连接"] D["全球学生创业大赛与跨国协作"] E["个人 AI 智能体实践课程"] end Challenge -->|"使传统单向讲授贬值"| Solution ``` --- ### [03:53](https://www.youtube.com/watch?v=oMgbGd95lSI&t=233s) Professor Labs 的愿景与布鲁姆教育目标分类法 回到教育痛点本身，Mike 回顾了他们在上一届大会上提出的核心构想：如何利用生成式 AI 与游戏化机制解决高等教育中三大根深蒂固的问题： 1. **千篇一律的静态课程（Static One-Size-Fits-All Curriculum）**：无法满足不同学生的个体节奏。 2. **学生学习动力的不均衡（Uneven Student Motivation）**：班级里通常有三分之一的学生一点就通、表现优异；三分之一完全缺乏动力；剩下的三分之一则处于中间地带，极度依赖外界激励。 3. **超个性化辅导的缺失（Lack of Hyper-Personalization）**。 团队基于经典的“布鲁姆教育目标分类法”（Bloom's Taxonomy）构建了底层教学逻辑。在常规教学中，前半学期通常聚焦基础认知（死记硬背、定义与基础练习），后半学期则过渡到实际应用与高阶分析，并穿插课堂互动与参与度。随着大语言模型（LLM）能力的演进，底层“记忆/定义”（Remember）类任务在早期模型中即可轻松胜任；而随着 GPT-4 及后续模型的迭代，AI 在“应用”（Apply）与“分析”（Analyze）等高阶认知层面上展现出惊人的生成能力。 Mike 特别指出，如果一家平台型创业公司不是 OpenAI 或 Google，就绝不应该在底层基础模型（LLM）层面盲目硬碰——那需要“每隔一两个月就烧掉 2500 万美元”的庞大资本；真正的核心竞争力在于基础模型之上的**垂直应用与教学框架层**。 --- ### [06:28](https://www.youtube.com/watch?v=oMgbGd95lSI&t=388s) AI 智能体架构：Knowledge AI、Student AI 与 Tutor AI 为了将愿景落地，Professor Labs 设计了一套四位一体的 AI 智能体系统架构，由多个协同工作的 AI 模块组成： ```mermaid flowchart LR accTitle: Professor Labs AI Agent Architecture subgraph ProfessorSide["教授知识资产"] K["Knowledge AI<br/>(教材/论文/最新PPT/课程IP)"] end subgraph StudentSide["学生数字画像与学习"] S["Student AI<br/>(知识图谱掌握度/数字孪生)"] T["Tutor AI<br/>(按学习风格无限自适应讲解)"] G["游戏化与性格驱动 (Voice)<br/>(微习惯记忆/打卡激励)"] end K -->|"注入课程知识"| T S -->|"反馈掌握短板"| T T -->|"自适应教学互动"| G ``` - **Knowledge AI（知识智能体）**：扮演教授的“数字大脑”与知识库。教授只需上传指定教材、最新发表的 4 篇论文或是刚刚优化的一套 PPT，所有的知识产权（IP）与前沿内容即被完整索引。 - **Student AI（学生智能体 / 数字孪生）**：精准追踪并量化每个学生的学习进度与知识掌握画像。例如记录学生在第四章掌握度仅有 20%，而在第五章达到 90%，从而动态映射出学生距离学期或每周教学目标的差距。 - **Tutor AI（导师智能体）**：充当个性化私人助教。针对传统课堂上学生对教授产生“听父母唠叨般的倦怠感”（如请客座讲师讲相同内容学生却觉得无比新鲜），Tutor AI 可以根据学生的偏好，自适应切换为视觉化、音频化、或“向五年级小学生解释”的口吻，无限生成定制化的讲解内容。 - **游戏化机制与人设（Gamification & Voice）**：契合年轻一代学生“手机不离手”的操作习惯（例如调侃现代年轻人比划“打电话”的手势早已变成盯着平整的手机屏幕），将枯燥的机械记忆转化为类似 Duolingo 般具有鲜明人设（甚至带点幽默式压迫感）的微习惯打卡机制。 --- ### [09:56](https://www.youtube.com/watch?v=oMgbGd95lSI&t=596s) 线下课堂的 AI 辅助：无限自适应评估与教育平权 在整体教学方法论上，Professor Labs 并不是要取代线下教学，而是作为**实体课堂教学的强力支撑系统**。 传统教育中存在显著的资源壁垒：富裕家庭的学生从小就能享受每小时 100 至 200 美元的一对一私人家教，而许多普通家庭的学生则缺乏针对性的辅导支持。平台的目标正是借助生成式 AI 为每位学生配备随时待命的顶级“个人导师”，敏锐洞察其薄弱环节并按其习惯进行互动沟通。 在考核评估方面，系统将课程内容系统化拆分为不同的大幕（Acts）与讲座单元（Lectures），并支持**持续且无限的自适应评估（Continuous Infinite Assessments）**。学生可以针对自己的薄弱章节无限次生成全新测试题进行针对性攻克，而教授则能从中一键抽样生成正式的期中考试卷，从而实现“练考同源但千人千面”的闭环学习流程。 ### [11:25](https://www.youtube.com/watch?v=oMgbGd95lSI&t=685s) 多模型自适应表达与全天候“私人导师” 团队在构建教学体验时，一直尝试根据学生的具体学习风格量身定制讲解方式。Mike Lee 谈到他们最近进行过的一项技术尝试：当时他们看到演示 demo 后非常兴奋，试图利用更先进的模型（如 GPT-5 早期预览技术）来生成动态视觉解释，比如创建一个展示风速与风力涡轮机运作机制的交互式 SVG，但实际效果并不理想，“根本行不通（doesn't work）”。尽管如此，这种尝试确立了一个核心设计理念：底层核心知识库（Knowledge AI）保持完全一致，系统按需灵活调用最合适的模型来呈现内容——若需要精美的视觉图表或生动语音，就接入最擅长该领域的模型，使知识的表达维度更加丰富。 在课堂互动与游戏化方面，系统内置的小游戏大多是轻量级的快速问答（Quick Q&A）。Mike Lee 打算将这些小游戏作为课堂参与度成绩（participation grade）的考核标准，例如要求学生每学期必须完成指定局数的游戏，但出乎意料的是，这种机制实际上“相当让人上瘾（pretty addictive）”。在学生端功能中，除了根据个人特征量身定制的“私人导师（Private Tutor）”外，还有“考核我（Assess Me）”功能，用于预测和演练可能出现的考试题目。学生可以全天候 24/7 随时提问，完全不必担心助教（TA）不回讨论区消息。Mike Lee 笑着调侃，学生们的作息向来颠倒怪异，总爱在半夜提问，“但他们绝不可能在早上 8 点找你，所以我故意把所有 Office Hours 都定在早上 8 点——想见我就得拿出真正的诚意”。 - **自适应表现层**：核心知识库保持固定，利用不同模型的专长动态渲染交互式图表、语音合成或文字解答。 - **24/7 学习陪伴**：弥补助教回帖延迟与学生深夜学习需求之间的断层。 --- ### [13:08](https://www.youtube.com/watch?v=oMgbGd95lSI&t=788s) 教授工作流与教材商业模式的重构 回顾最初的产品设想，整体工作流分为明确的几个阶段：教授上传课件 PPT、教材书籍或学术文章，Knowledge AI 进行深度解析并自动转化为一份完整的教学大纲（Syllabus）；随后教授作为掌控者进行编辑微调，剔除不需要的话题或补充新章节；紧接着系统自动生成配套的课堂小游戏，并在第七步激活高度个性化的 Tutor AI。Mike Lee 强调，这是一个真正“由教授打造、为教授服务（for professors by professors）”的工具。 ```mermaid flowchart LR accTitle: Professor Labs Workflow accDescription: 工作流从教授上传材料开始，经历知识解析、大纲生成、人工微调，直至生成交互游戏与个性化导师。 subgraph Input["输入阶段"] A["上传教学资料<br/>(教材/PPT/论文)"] end subgraph Process["AI 解析与生成"] B["Knowledge AI 解析"] --> C["自动生成 15 周大纲"] C --> D["教授人工微调审阅"] end subgraph Output["交付体验"] D --> E["生成互动小游戏"] D --> F["激活超个性化 Tutor AI"] end A --> B ``` 然而，在推进过程中团队发现了更深层的现实阻碍：教授群体普遍缺乏将优质课件转化为数字化内容的动力。传统模式下，即便教授费心编写了高价教材，学生往往只需花几百美元买一本，随后便被翻印成 PDF 免费流传。硅谷的传统软件工具无法改变这种生态，但大学课堂拥有天然的制度壁垒——“只要教授指定某项材料为必读，它就是必读（required reading）”。因此，Professor Labs 正在将纯粹的“备课工具”升级为“互动课程订阅平台”：教授基于自身知识资产生成的 Knowledge AI，能够以每生每学期 10 到 20 美元的许可费形式授权给学生。这不再是卖软件工具，而是为教授独一无二的智力成果定价。 与此同时，团队高度关注 AI 技术的普惠与民主化（Democratization of AI）。Mike Lee 提到他在前一天论坛上遇到的一位金融学女教授，虽然背景是商科，却展现出了顶级工程师级别的极客应用水平；但现实中绝大多数教师并非技术专家。不仅美国本土只有两到三成教师能熟练驾驭这些前沿工具，在人均 GDP 较低的发展中国家，技术与成本鸿沟更为严峻。团队提出了基于购买力平价的动态定价模型：例如在美国基准定价为 50 美元的课程知识库，在印度等发展中国家可折算为仅需 1.59 美元即可使用。 --- ### [16:18](https://www.youtube.com/watch?v=oMgbGd95lSI&t=978s) 平台解耦架构与三种部署模式 在技术架构层面，Professor Labs 的核心设计原则是“底层解耦”。基础大语言模型每隔一两个月就会迭代一次，因此系统的核心知识产权和专有数据必须牢固沉淀在 Professor Labs 应用层内。其技术栈完全基于开源构建，具备平台无关性（Platform-agnostic），可根据任务成本与性能需求无缝切换模型。例如在需要生成复杂的交互式可视化图表时调用能力更强、单价较高的前沿模型（如 GPT-5 预览类模型），而在处理日常问答时则回退到性价比更高的 GPT-4 级别模型。 ```mermaid flowchart TD accTitle: Professor Labs Architecture and Deployment Models accDescription: 展示应用层核心与底层模型解耦架构，以及本地、混合、云端三种部署路径。 subgraph AppLayer["Professor Labs 核心层 (专有 IP & 数据)"] Core["课程结构化引擎 + 知识库编排"] end subgraph Models["动态调用模型池 (成本与能力自适应)"] M1["高性能前沿模型 (复杂图表/深度推理)"] M2["高性价比模型 (日常问答/大纲提取)"] end subgraph Deploy["三种部署形态"] D1["纯本地部署 (离线/高保密/需强劲 GPU)"] D2["混合部署 (本地存储 + 外部商业 API)"] D3["纯云端部署 (推荐形态/多端 24/7 访问/免训练承诺)"] end Core --> Models Core -.-> Deploy ``` 团队系统性测试了三种不同的落地部署方案，各自对应不同的数据主权与硬件门槛： - **纯本地部署（Local）**：所有数据和大模型推理均运行在教师个人的笔记本或工作站上，支持在完全断网环境下根据个人教材生成课程体系。这种模式数据安全性最高，但对本地 GPU 性能有极高要求，且本地小模型的推理表现与顶级云端模型相比存在明显差距。 - **混合部署（Hybrid，本次现场演示所用方案）**：Professor Labs 系统及全部核心教学数据保存在本地设备上，仅将文本处理请求通过 API 抛给 OpenAI 等云端服务商进行算力处理。该方案兼顾了一定的隐私性与算力，但缺陷是学生无法直接远程访问本地数据库，无法实现深夜随问随答。 - **纯云端部署（Cloud）**：团队认为的最理想商业落地形态。既能保证全班学生 24 小时随地访问，又能依托企业级协议确保所有上传的专有课程数据绝对不被大模型厂商用于模型二次训练。 --- ### [19:12](https://www.youtube.com/watch?v=oMgbGd95lSI&t=1152s) 现场实机演示：15周大纲生成与即时知识挂载 Manuela Berrio 与 Mike Lee 开启现场实机演示，生动展示连通学生与教师的“双向教育车道”。首先是教师端的课程规划流程：系统默认规划为 15 周的课程架构。教师首先指定期中考试的时间节点（演示中设在学期中段的第 7 周），随后上传了 Mike Lee 自己的著作 PDF 作为核心教学蓝本，点击生成按钮后，系统迅速抽取出完整的学期教学大纲。 在系统界面中，最左侧清晰罗列出每周的课程日程（按每周一、周三上课进行排期），中间展示大模型从教材 PDF 中切片提取出的核心课题；若教师觉得某些关键概念被遗漏，界面还贴心提供了“备选建议课题（Additional Topics）”供随时替换。紧邻其后的是 Knowledge AI 模块，允许教师在大纲基础上为特定周次追加专项 PPT 幻灯片、补充阅读或授课备忘录。 为了向现场观众验证 Knowledge AI 的动态外挂与精准检索能力，两位演讲者邀请现场观众提供一条原版教材中绝不可能存在的个人冷知识。观众席上一位名叫 Alex 的参会者自告奋勇：“1980 年圣海伦斯火山（Mount St. Helens）爆发时，我正好五岁。” Mike Lee 随即在后台创建一个文本文件，输入 Alex 的这条个人信息并保存上传，试图演示 Tutor AI 如何在毫秒间完成对新增非结构化知识的向量吸收与问答响应： > **Mike Lee**：“在圣海伦斯火山爆发时，Alex 当时几岁？” 在现场演示过程中，由于操作节奏匆忙与文件本地保存路径切换出现小插曲，系统在首次上传检索时未能即刻抓取到这条数据，Mike Lee 坦率而幽默地化解了这一现场技术抖动，并迅速带领观众进入下一个更具价值的核心功能展示环节。 --- ### [23:38](https://www.youtube.com/watch?v=oMgbGd95lSI&t=1418s) 智能生成主题讲义与课件大纲 在解决知识挂载的同时，Manuela Berrio 演示了系统面向教师日常备课痛点的另一项核心输出能力——自动生成“主题讲义大纲（Topic Notes）”。该功能旨在将大纲中的抽象主题迅速转化为可落地的授课材料。 - **幻灯片逐页规划**：系统自动构建匹配 Google Slides 或 PowerPoint 的演示架构，明确规定每一页幻灯片应当呈现的关键要点与视觉内容。 - **逐页讲者备注（Speaker Notes）**：为教师或代课助教生成详尽的口头讲解提示词，帮助把控课堂讲授节奏。 - **深度知识点拆解**：针对选定的单个核心课程主题，系统会自动输出一份包含 20 个细分知识维度的结构化解析方案，帮助教授省去从零整理教案的重复劳动。 ### [24:11](https://www.youtube.com/watch?v=oMgbGd95lSI&t=1451s) 多语言支持与未展示功能展望 在演示接近尾声时，Manuela Berrio 补充指出了界面中一个此前未及详述的紫色功能按钮，并分享了自己的个人背景。她提到自己来自哥伦比亚并在那里生活到 17 岁，因此英语并不是她的母语，这也正是多语言学习支持对国际学生至关重要的原因。由于现场演讲时间所限，他们必须进行收尾，但热情邀请有兴趣深入了解的与会观众在演讲结束后移步场边，继续观看更详尽的线下实时演示。 Mike Lee 随后对本场因时间关系未能完整演示的核心功能进行了补充概括。他特别强调了平台出色的多语言能力，支持包括韩语、德语、西班牙语在内的多种语言无缝切换，其翻译水准依托于当前顶尖的大语言模型技术。此外，平台还具备基于课程内容自动生成小测验的小游戏（Mini-games）模块，以及功能强大的期中考试生成器（Generate Midterm），能够一键生成包含 30 至 40 道题目的完整考卷，并全面支持多选题与实际应用题等多种题型。 - **多语言适配**：支持韩语、德语等跨语言无缝转换，借助前沿 LLM 提供高质量学术翻译，帮助非母语学生消除学习障碍。 - **互动测验小游戏**：利用 Mini-games 动态生成配套小测验，实现寓教于乐的游戏化自测体验。 - **期中考题自动生成**：系统可一键生成 30-40 道题规模的期中测试卷，涵盖客观选择题（Multiple Choice）与深度应用题（Application Questions）。 ## Notable quotes > [03:53](https://www.youtube.com/watch?v=oMgbGd95lSI&t=233s) 每隔一两个月就烧掉 2500 万美元 — Mike Lee > [11:25](https://www.youtube.com/watch?v=oMgbGd95lSI&t=685s) doesn't work — Mike Lee > [11:25](https://www.youtube.com/watch?v=oMgbGd95lSI&t=685s) pretty addictive — Mike Lee > [11:25](https://www.youtube.com/watch?v=oMgbGd95lSI&t=685s) 但他们绝不可能在早上 8 点找你，所以我故意把所有 Office Hours 都定在早上 8 点——想见我就得拿出真正的诚意 — Mike Lee > [13:08](https://www.youtube.com/watch?v=oMgbGd95lSI&t=788s) for professors by professors — Mike Lee > [13:08](https://www.youtube.com/watch?v=oMgbGd95lSI&t=788s) 只要教授指定某项材料为必读，它就是必读 — Mike Lee > [19:12](https://www.youtube.com/watch?v=oMgbGd95lSI&t=1152s) 1980 年圣海伦斯火山爆发时，我正好五岁。 — Alex > [19:12](https://www.youtube.com/watch?v=oMgbGd95lSI&t=1152s) 在圣海伦斯火山爆发时，Alex 当时几岁？ — Mike Lee ## Mentioned Mike Lee, Manuela Berrio, University of Southern California (USC), Professor Labs, Netscape, Golden Apple Award, Global Student Startup Competition, UC Berkeley, Harvard University, Cornell University, University of Oxford, Aalto University, European Innovation Academy, Bloom's Taxonomy, OpenAI, Google, GPT-4, GPT-5, Knowledge AI, Student AI, Tutor AI, Duolingo, Google Slides, PowerPoint ## Video description Mike Lee, Senior Lecturer of Digital Entrepreneurship & Advanced Computing, University of Southern California Manuela Berrio, Undergraduate Research Assistant, University of Southern California Abstracts At AI4 2024, we introduced Professor Labs as an academic project aimed at transforming higher education through generative AI and gamified learning. One year later, we return with a working prototype and a live demonstration. In this session, Mike Lee (Senior Lecturer of Digital Entrepreneurship & Advanced Computing, University of Southern California) and Manuela Berrio (Computer Science & Business Administration major, University of Southern California) showcase how Professor Labs enables professors to transform traditional lectures into interactive, AI-powered learning quests. Built by and for educators, the platform leverages custom large language models trained on course-specific content to deliver hyper-personalized tutoring, continuous assessment, and student-led gameplay. This live demo will highlight: - The Professor Use Case: Uploading course content, generating assessments, and guiding pedagogy - The Tutor AI: Real-time, personalized feedback and support tailored to individual students Join us to see how academic research is translating vision into practice—and shaping the future of teaching and learning. [Source](https://www.youtube.com/watch?v=oMgbGd95lSI)

Source quotes

[3:53] Mike Lee: 每隔一两个月就烧掉 2500 万美元 [11:25] Mike Lee: doesn't work [11:25] Mike Lee: pretty addictive [11:25] Mike Lee: 但他们绝不可能在早上 8 点找你，所以我故意把所有 Office Hours 都定在早上 8 点——想见我就得拿出真正的诚意 [13:08] Mike Lee: for professors by professors [13:08] Mike Lee: 只要教授指定某项材料为必读，它就是必读 [19:12] Alex: 1980 年圣海伦斯火山爆发时，我正好五岁。 [19:12] Mike Lee: 在圣海伦斯火山爆发时，Alex 当时几岁？

Source quotes

[0:04] Mike Lee: Thank God... it's been like five years since a single student knew what that was. So I really appreciate all of you. [3:53] Mike Lee: how the hell did they get into our university [6:28] Mike Lee: as if they're a fifth grader [10:12] Mike Lee: At any time, they could ask this 24/7. Doesn't matter if your TAs don't respond to discussion boards... because students have wacky schedules—midnight, whichever time, etc. Never at 8:00 AM. That's why I make all my office hours 8:00 AM. You must really want to see me. [21:17] Alex: When Mount St. Helens blew in 1980, I was five years old.

Original transcript

[0:04] All right, welcome, welcome. So, my name [0:07] is Mike Lee. I'm a senior lecture of [0:10] digital entrepreneurship and applied [0:12] computing at the University of Southern [0:13] California. Um, and this is my wonderful [0:18] student. [0:18] >> My name is Manila Rio. I'm a rising [0:21] junior at USC studying computer science [0:23] and business administration with an [0:25] specialization in AI applications. and [0:28] we're really excited to be here and for [0:30] you to join us today. [0:32] >> Yeah, she's amazing. She's one of my top [0:34] students and uh she's doing research on [0:37] this area with me and she's also going [0:38] to be a TA for me on building personal [0:41] agents. So, that's what we have all of [0:43] our students do now. It's like build [0:45] personal AI agents, LLM. So, I feel like [0:48] they're all going to be AI natives when [0:51] they graduate. So, I'm pretty excited. [0:53] Um, hold on a sec. So a little bit more [0:57] uh about me. So my background is over 20 [1:01] years of experience being a CTO at uh uh [1:05] media companies and um d startups etc. I [1:10] was also over at uh do you guys remember [1:12] Netscape? [1:13] >> Thank god [1:16] my students it's been like five years [1:18] since a single student knew what that [1:20] was. So I really appreciate all of you. [1:23] Um I was a technology evangelist there. [1:25] I was actually, you know, here's a [1:27] technology, tell the world about it, [1:29] things like that. I spent a decade at [1:31] the studios as a vice president of [1:32] technology. Um, and did a lot of, uh, [1:37] venture building with the latest [1:40] technologies for a while. Then I decided [1:42] to take [1:44] one of my salary and then start working [1:46] at USC teaching. Who are the teachers [1:48] out there? Raise your hand. Yes, we [1:51] could make five times more in the [1:53] industry, but we we actually like our [1:56] students for some reason. I'm not quite [1:58] sure why, but so I teach over at uh USC. [2:03] I in the digital entrepreneurship [2:04] applied computing area. I'm a faculty [2:06] adviser to a lot of the innovation uh [2:08] organizations. I'm teaching or building [2:11] out the uh uh personal AI agent uh [2:14] courses there. At this point, I actually [2:17] won a Golden Apple award. I never win [2:19] anything that is faculty voted but I win [2:22] things if it's student voted. So that's [2:24] one of the things that I won there. Um [2:26] I'm also founder of global student [2:28] startup uh competition. So we basically [2:32] bring over 50 students from [2:34] international [2:35] fly them all fully funded to Korea to [2:38] compete with 50 students in Korea. We [2:42] also fly in 20 mentors and faculty to [2:45] mentor them. Right? Because the whole [2:48] whole idea is I like for me anyways with [2:50] AI and YouTube and all of those things. [2:53] Um [2:55] what's the value of university right? I [2:58] feel that it's really that experiential [3:00] the things we can't replicate is just [3:02] you and I meeting uh right here. So I'm [3:05] my mission is to put that together. So [3:07] we got fortunate enough to get funding [3:09] for this. So this year we're going to [3:11] have USC, Cal, Harvard, Cornell, um [3:16] Oxford and Alto, which is a Finnish [3:18] university coming and also [3:19] representatives from Europe and all [3:21] that. Um I'm also a head of mentors over [3:24] at European Innovation uh Academy. I [3:27] just came back from there like a day [3:30] ago. I came home, hugged my dog for a [3:33] day, and then got in the car, came over [3:35] here straight. Um, but there's about 300 [3:38] students from all over the world there [3:39] as well. And then one of the things that [3:41] came out was all of these students know [3:43] how to build agents and quickly turn [3:46] things around. So it's it's amazing how [3:49] fast everything is uh uh turning around. [3:53] So we actually presented here last year [3:57] uh it was on hyperpersonalized [3:59] uh uh learning through AI and [4:01] gamification. So we set out a vision to [4:04] solve certain core problems which is you [4:07] know static one-sizefits-all curriculum [4:09] uneven student motivation you know you [4:11] know what our students are like we have [4:13] one-third amazing you give them one word [4:16] and they just do everything amazing we [4:18] have this other third it doesn't matter [4:20] what the hell you do you wonder how the [4:22] hell did they get into our university [4:24] right and there's the middle which is [4:26] depending on how well you motivate them [4:30] right so there is a a uneven motivation [4:33] on that side and like the previous [4:35] panelist said it's hyperpersonalization [4:39] for each student. So we decided to [4:41] tackle this problem and then we we came [4:43] up with the vision for professor labs. [4:46] Um so the you know there's the blooms [4:50] taxonomy which is why did we pick this [4:52] one? because we had to pick one that's [4:55] the reason why but I we translated that [4:58] into [5:00] the way we look at it at least in my [5:01] classes I always think about what are [5:04] the fundamentals first half of semester [5:05] is fundamentals maybe ro memorizations [5:08] maybe exercises etc then there's a [5:11] second part which is can you apply what [5:13] you have learned right and I also [5:15] sprinkle in a little bit of uh [5:16] participation in class exercises you [5:19] came to class you engaged etc right so [5:22] But we put that all together on top of [5:24] Bloom's taxonomy and we add a gaming [5:27] element to this. What's interesting is [5:30] see this over here. Well, the Bloom's [5:33] taxonomy right here. [5:37] As the LLMs are improving, we're able to [5:41] get the same content and create [5:44] exercises and study guides in an [5:47] infinite manner with each progression. [5:50] Right? So at the bottom it says remember [5:53] that's just definitions we could do that [5:56] since the first version but when we [5:58] start going up higher how to apply how [6:00] to analyze well the L&M's are getting [6:03] smarter right so if you're a platform [6:07] company that's not named OpenAI or [6:10] Google you're not going to compete in [6:12] the LM space what you're really [6:14] competing in is the layer above that [6:17] right so if you hear anybody trying to [6:18] compete down here just like B they're [6:20] full of right? Or they have $25 [6:23] million burning in their pockets every [6:25] month or two, right? So, as part of [6:28] this, we generated a framework for our [6:30] AI bots. In the center of this is [6:34] knowledge AI. So, for us in our context, [6:37] what it really is is like I have a book. [6:40] I load it into the knowledge AI. Now, it [6:44] contains the knowledge. Maybe this year [6:47] there's four new articles I like. I load [6:49] that into knowledge AI. So this [6:51] represents your brain, right? And we [6:54] have a demo that'll show uh this portion [6:56] as well because it might be that you [6:58] just created a a better PowerPoint for [7:01] your new lecture. Let's load that in [7:03] there for that topic, right? So this is [7:06] all of your intellectual property and [7:09] what you do. Student AI is this is you [7:13] know the digital twin. It just means [7:15] what is the progress of the student. [7:17] they may be uh a test they out of 100 [7:21] they got section four 20% section five [7:24] 90%. So every student has different [7:28] things except for the top student. I [7:29] think 100% they're perfect and we just [7:31] go like yes thank you for coming. Uh but [7:36] this represents where are they relative [7:38] to the knowledge that they should earn [7:41] by the end of the course right even end [7:43] of this week. Um, additionally, we have [7:45] tutor AI. So now, how do we explain this [7:48] to our students, right? Yes, there's the [7:52] book reading. Here's a lecture. How [7:54] brilliant we are. But you know how we [7:55] become like parents to them. Have you [7:58] noticed that? Have you ever brought in a [7:59] guest lecture? They say the exact same [8:02] stuff that you did except they're like, [8:03] "Oh my god, that's amazing." Right? So [8:07] if you haven't taught, we do become [8:10] parents and it goes that thing. Right? [8:12] So what this does, depending on their [8:14] learning style, visual, audio, or maybe [8:17] they just need it explained to as if [8:20] they're a fifth grader, all of those [8:23] things are infinitely generated. [8:26] And that generation gets better with [8:28] GPD4, GPT5. [8:31] It's incredible. [8:34] Yeah. And we'll give demo demo of that [8:35] as well. Uh there is a gaming aspect. So [8:39] we translate this knowledge into a [8:40] format that is familiar to our students [8:44] which is staring at their phones, right? [8:47] Oh my god. So who knows cats? [8:51] Yes, I love you, man. [8:53] So catsai, they have this you it's a a [8:56] K-pop group made of international [8:58] people. So they have a game they're [8:59] playing. They have two girls up there. [9:01] They're like they say a word and you [9:03] have to act it out. So they said phone. [9:05] One girl goes like this. I would do [9:08] this, right? Do you know what the other [9:09] girl did? [9:11] Like that she has it. She's young. You [9:14] know how crazy that is? That a phone to [9:16] them is this, right? [9:19] So, it just means that for them when [9:23] they're waiting, we could actually have [9:24] them do the rope memorization part and [9:27] it's just like a little habit type of [9:28] thing. So, we could convert this to any [9:30] type of gamification we want. Voice just [9:32] means um like the personality. Who uses [9:35] Dualingo? [9:38] Have you ever had it remind you in a [9:40] very aggressive way? There's memes about [9:42] it where they're threatened to kill you [9:44] like follow you if you don't do it [9:46] right. But there's a personality to it. [9:48] But this is what our students or [9:50] especially our undergraduate students [9:52] really know about, right? So we break up [9:56] courses. You load up the content breaks [9:57] it up to acts. X just means like the way [9:59] we would separate in terms of midterms, [10:02] right? And we have the lectures [10:03] underneath just like we would do, right? [10:06] So [10:08] just I'm going to go through quickly so [10:09] we could get to the demo. [10:12] So in terms of the way we look at the [10:13] learning uh methodology for us is this [10:16] is to support in-person learning. So [10:19] obviously we had the prerequisites and [10:21] you know those tests that we could do. [10:23] We do our Inerson lecture and then this [10:27] is where they have a personal tutor. [10:30] So if you're rich, some of our students [10:33] are rich. They had $100 an hour, $200 an [10:37] hour tutors since they're five years old [10:41] versus [10:45] sometimes the parents are so busy that [10:46] they have no idea what the kids are [10:47] doing, right? We want to replicate that [10:51] feeling of a personal tutor for that [10:53] person. Where are your weaknesses? How [10:55] do you communicate? Right? So for me, [10:57] that's the power of of AI, right? Is [11:00] really there. Um, and here's the part [11:04] where's this continuous infinite [11:05] assessments, they could always continue [11:08] to test on that one section. Why? [11:10] Because we could generate infinite [11:12] number of questions, right? one of the [11:14] generations happens to be the midterm [11:17] that we would test on, but the students [11:19] could run these for infinite, right? [11:22] Additionally, it's not only that we [11:25] could explain it to them in their [11:26] specific learning style. So, one of the [11:29] things we tried and failed like she [11:31] tried and failed recently is is the fact [11:34] that we use GPD5 for their uh visual [11:37] explanations. I got excited because I [11:39] saw the demo. Create a interactive SUV [11:43] of this content. You know, they're [11:46] showing the wind speed, turbine, all [11:48] that doesn't work, right? [11:51] So, I'm sure it'll get there, etc. But [11:54] basically, what it means is we use [11:56] whichever model that makes sense to add [11:59] that visualization or add that speech, [12:02] right? So as it turns quickly, your [12:05] knowledge AI is exactly the same. [12:08] It just the way it expresses itself is a [12:10] lot better. [12:12] Uh in terms of gaming, our mini games [12:14] are uh just quick Q&A. Like for me, my [12:18] plan is to use this just for the [12:20] participation grade, right? That you [12:22] have to do X number of of games. But you [12:25] know what? This get thing gets pretty [12:27] addictive. [12:29] Okay. back and go back here. A couple [12:33] things, you know, where it says private [12:35] tutor, that's a personalized one based [12:36] on you. The other one is uh assess me. [12:39] What are the test questions that's going [12:41] to come up, right? At any time, they [12:44] could ask this 247. Doesn't matter if [12:46] your TAs don't respond to discussion [12:49] boards. I have that a lot, right? Um or [12:54] even, you know, because students are [12:56] wacky schedules, right? midnight, you [12:59] know, whichever time, etc. Never. Never [13:01] at 8:00 a.m. That's why I make all my [13:03] office hours 8 a.m. You must really want [13:06] to see me. [13:08] So, let me just go back here. Okay. So, [13:11] the use case that we came up last time [13:14] was professor uploads your materials, [13:16] PowerPoints, books, articles, whatever [13:19] one. The knowledge analyzes it and [13:22] converts it into a syllabus, right? And [13:25] then professor edits it. meaning oh I [13:27] don't want this topic let's add another [13:29] topic all of those so it's like a tool [13:31] for you right it's is a tool uh for [13:34] professors by professors [13:36] um then it could generate the gaming of [13:39] that the other part is step seven is [13:41] really where it's very powerful is that [13:43] tutor AI now becomes hyperpersonalized [13:47] right [13:50] so what we learned sensing there's [13:52] actually additional problems [13:56] One, the professors now have zero [13:58] incentive to produce content and make it [14:01] available. Yes, maybe your tenure and [14:04] all those things, but if there's any [14:06] money part, forget about it, right? But [14:09] let's see what happens here. You write [14:11] the book, student has pay hundreds of [14:13] dollars for the book. So what so what [14:16] what we end up doing and not me I did it [14:20] for professor Kali's book because I need [14:21] it for the demo but uh they put into PDF [14:24] and it's free for everybody is really [14:26] what happens right so now the model that [14:31] we want to do is not just a tool for [14:32] just creating your your class but now [14:35] it's going to up uh it's going to turn [14:37] into curriculum and now it's [14:40] subscription [14:42] right think about it Silicon Valley [14:43] cannot replicate that which is if I say [14:46] this is a required reading it is [14:48] required reading [14:50] right so now imagine your knowledge AI [14:53] that you created is licensed for 10 $20 [14:56] per student per semester it's not the [14:59] tool it isn't the fact they have access [15:01] to it it's your knowledge [15:04] right so I think this is you know a lot [15:07] of the those uh companies uh do that as [15:10] well right in terms of let's make an [15:13] online book but this is truly [15:14] interactive. [15:16] Uh the other part we learned was that uh [15:19] democratization of AI. I met a professor [15:22] yesterday. She was on a panel. She's [15:24] amazing. She's a super she's a finance [15:26] person. I'm like you're an engineer but [15:28] such a powerful super user. But guess [15:31] what? Not [15:33] everybody geeks out that way, right? And [15:37] we might be at US maybe 20 30 50%. But [15:41] if we go to other countries that has is [15:45] lower GDP etc. it doesn't like that. So [15:48] how do we democratize this right? How do [15:51] we encapsulate all that where it makes [15:52] it easy and then how do we even have [15:56] this available for the students? And I [15:57] just put this as an example we baseline [16:00] it at uh US let's say $50. Um and then [16:04] if we have this in India based on GDP [16:06] it's only a$159 [16:09] right? So that sort of is is what we're [16:12] saying about democratization [16:14] part of it. So I want to show you some [16:16] of this. [16:18] So when we were building this [16:23] we realized that [16:25] the bottom layer is going to improve [16:28] every month, every two months. So the [16:30] real intellectual property is within the [16:34] professor labs portion of it, right? And [16:37] all your data is stored there. [16:40] AI is used for there is pre-trained and [16:42] in and it could process things. All your [16:44] AI is stored there. The tech stack is [16:47] open source. Uh so it's AI platform [16:50] agnostic. So we could use a local on [16:52] your laptop or or or PC or we could use [16:57] a cloud. Well, the difference is the [16:58] power, [17:00] right? Um [17:03] and it improves based on the industry. [17:06] any single thing pops up we'll we'll [17:08] improve with it. So in terms of the AI [17:10] platform improvement over time cost but [17:13] it's very cost proerative to duplicate. [17:15] So we make sure there's a clear [17:17] separation and we could use any model [17:21] for our purposes. So if we're doing the [17:23] visualization portion of it interactive [17:25] visualization GPD5 because it's a little [17:27] bit more expensive everything else GPD4 [17:29] is better for us. Does that make sense? [17:33] So we could act very very very fast in [17:36] this scenario. [17:39] Deployment scenarios, we've tested out [17:41] three of them. Uh the local one is [17:44] everything's running on your computer. [17:46] Incredible. You could do it for course [17:48] generation based on your books, but all [17:50] data is on your your computer. And guess [17:53] what? It even runs without without [17:54] internet, right? It's pretty crazy some [17:57] of the things that you can actually do [17:59] just specifically just on your computer [18:01] itself with those LLMs. Um but you need [18:04] a powerful GPU and AI just really can't [18:08] I mean versus GPD5 or four is very [18:12] difficult to do there. Um there's the [18:14] hybrid one which we'll demo today. So [18:18] this is where the professor labs is [18:19] local. So what this means that all the [18:22] data stays on your computer but we're [18:24] using the processing power of a open AI [18:28] or or whichever API that you want to [18:30] use. Okay. The problem with that is [18:32] there's no direct student access. So you [18:35] know that part about middle of the night [18:37] students have questions. Well that's not [18:40] available because it's actually stored [18:42] on your computer itself right. So the [18:45] way we're seeing it is we're cloud is is [18:48] probably the most ideal for us and we [18:50] use everything where none of the data is [18:53] used for training. Right? [18:57] Okay. So [18:59] see here just to set up the demo itself. [19:01] We're going to do parts of the use case [19:04] and it's going to be the knowledge AI [19:06] portion and then still AI. [19:09] Oh. [19:12] >> Perfect. So, demo time. We are thrilled [19:15] to show you our first version of [19:17] Professor Labs, and we really want to [19:20] highlight that two-way street of [19:22] education where we have students on one [19:24] end and professors on the other, which [19:27] is why you see Professor Lee and myself [19:30] working on this together. And so, at the [19:32] beginning, we'll work with the [19:34] professor. This will create a 15 course [19:37] plan. And we'll first ask the professor [19:41] what week he wants he or she wants the [19:45] midterm to be. So initially we'll put in [19:48] week seven which is ballpoint the middle [19:51] of the course and then professor will be [19:54] able to upload any book PDF presentation [19:59] to create and generate the syllabus. So [20:01] in this case we'll use professor Lee's [20:03] book. We'll upload it and we'll click [20:08] generate and it will generate the [20:09] course. And so we want to highlight two [20:12] main points. Number one, professor labs. [20:16] So hyperpersonalization for the [20:18] professor where we have knowledge AI [20:20] where they can add more content other [20:23] than their book. And then we also have [20:25] tutor AI at the far right corner which [20:28] is meant for students. But the for the [20:30] sake of this demo, we have all of it [20:32] right now. And as as as professor [20:34] mentioned, we want to move forward to [20:36] cloud-based, but for now, we'll show it [20:39] all here. So on the far left corner, [20:42] you'll see the weeks with the dates. So [20:44] for this course specifically, we have [20:46] classes every Monday and Wednesday. [20:49] And we'll see the lecture topics that [20:51] the LLM chunked and took out of the PDF. [20:55] And we also have additional topics in [20:57] case the professor deems another topic [21:00] more relevant that the ones the AI [21:02] chose. Then we have knowledge AI column [21:06] where we can add more. So we already [21:09] have the book but what if the professor [21:11] wants to add a specific slide for that [21:13] topic, specific notes, right? And to [21:17] test this out, we really want your [21:20] input. So if we have a volunteer that [21:24] maybe wants to share a personal fact or [21:26] something that professor Lee's book does [21:28] not have at all so we can test how this [21:32] works [21:33] >> like social security numbers made a name [21:37] >> but any fact [21:39] >> yes anything [21:41] >> you have a brave person. [21:44] >> Okay. [21:44] >> Yes. [21:46] >> Or about anything. [21:48] >> Yeah. About you. [21:49] >> So yourself. Oh, uh, [21:51] >> what's your name? [21:52] >> Alex. [21:52] >> Alex. [21:55] >> And, uh, let's see. I was, uh, when [21:58] Mount St. Helens blew [22:00] >> Yeah. [22:00] >> in 1980, [22:03] >> I was five years old. [22:08] >> He was five. [22:11] Okay. So, I'm just going to save this [22:13] here. Let me do this really quick. So, [22:15] first of all, I'm gonna ask Tutor AI. [22:18] Uh, [22:20] how old was Alex? [22:26] I'm sorry. Right. So, what we're going [22:28] to do is get this. Oh, do I need to save [22:31] that one? [22:31] >> Yes. [22:32] >> Okay. So, back here. [22:35] File. [22:37] Save as. [22:38] >> Yes. download ass [22:43] like that. [22:45] Okay. So now back here [22:49] upload. [22:52] Do you know where it is? [22:56] Right there. [23:01] All right. So we just uploaded it. So [23:02] let's ask it again. [23:05] How old was Alex? [23:16] How old was Alex? [23:19] Mount St. Helens, right? [23:25] Oh, no. [23:29] Is this the right one? Maybe. [23:34] Okay, I'll go to the next one. We're [23:36] going to try this in a sec. [23:38] So the other thing is it just has short [23:40] time right now. So what we're going to [23:42] do is generate the topic notes, [23:46] right? [23:48] >> So these topic notes are basically a [23:51] course or not a course, a topic outline [23:54] for the professor to be able to generate [23:56] a PowerPoint slide or Google Slides. It [23:59] has speaker notes. It has what each [24:02] slide should have and it has a [24:04] specifically 20 topic breakdown of that [24:07] specific topic. And [24:11] something I haven't mentioned is we see [24:13] there's also purple button that says [24:18] and I didn't mention but I'm also from [24:21] Colombia. I grew up and I lived there [24:23] until I was 17. So naturally English is [24:26] not my first. If you want actually we [24:28] have to wrap up but if you want to see [24:30] the demo will be over there with the [24:32] demo. [24:33] >> Um the things that we didn't show yet uh [24:37] what like she mentioned is the multil [24:40] language portion of it like Korean, [24:43] German and all those things and it it is [24:46] exactly the same in terms of the [24:47] translation with the best LLM [24:50] translations at this point. Uh the mini [24:52] games generates the quizzes and then the [24:57] generate midterm creates your 30 40 type [25:00] questions and then you could do both [25:02] multiple choice and application [25:04] questions as well.